The hard part of Islamic AI is not the model

The first version of Madeenan felt deceptively simple. Gather Islamic source material, connect it to a model, and ask for a cited answer. This is the version of the product that fits neatly on a whiteboard.

Then I started asking ordinary questions.

Ordinary questions quickly made the problem clearer. Some answers found one helpful source and then wandered. Some brought useful reports into the same space as weaker material without enough caution. Some familiar references were obvious to a human and still easy for software to mishandle.

The model was not the main problem. The model was doing what models do: turning whatever context we gave it into fluent prose. The harder question was whether the context deserved that fluency.

In Islamic AI, answer quality begins before the model writes a single word.

Retrieval is a product surface

It is tempting to think of retrieval as plumbing. Users never see the machinery underneath. They care whether the answer feels careful and whether they can inspect the source behind it.

But retrieval is exactly where many product decisions hide. Which source should be shown first? How much uncertainty should be visible? When should the product show raw sources instead of a generated explanation? Those are not back-office questions.

These are not just search-quality choices. They shape the answer before the answer exists.

The source document matters

One of the biggest lessons was that a source is not just a block of text. A verse, a hadith, a commentary note, and a remembrance entry each carry their own context. If the product treats them as interchangeable prose, the answer becomes easier to misread.

When source context is clear, the answer has fewer excuses to improvise. The user also gets a better chance to see what kind of material is being used before trusting the explanation.

The answer has to know when to stop

Islamic questions often invite overreach. A user asks, "Can I do this?" and the product wants to be helpful. But if the retrieved sources establish only part of the issue, the answer should say so. It should explain the sources, name the boundary, and tell the user to ask a qualified scholar for a practical ruling.

This restraint is not a weakness in the product. It is part of the product.

What changed after February

The work became less about finding a magic model and more about building a better source path. References needed to be easier to inspect. Different kinds of Islamic material needed to be handled with their own caution. The answer needed less grandiosity and more discipline.

A better model still helps. It can write more beautifully, follow instructions more reliably, and handle nuance with more grace. But the model cannot rescue a bad source set. It can only make the failure sound smoother.